The question this week was whether the AI boom is ending. I read every hyperscaler’s earnings, and none of them cut its plans. Microsoft expects to spend $170 billion this year, Alphabet $195 billion to $205 billion, Amazon $220 billion, and Meta up to $145 billion. Microsoft’s trim from $175 billion happened because its GPU suppliers are late, and its finance chief said the spending will speed up again in the first half of 2027.
Demand is not fading either, judging by what the companies reported. Alphabet’s cloud business more than doubled in a year. Anthropic’s filing shows revenue up about fourteen times on the year before. CNBC notes that Claude has at times been unavailable this year because demand outran capacity.
What did change is where the money comes from. For years these companies paid for data centres out of the profits of search, advertising, shopping and software. This quarter three of the four big spenders had negative free cash flow, meaning they spent more than their businesses brought in. Amazon was out $42.4 billion, Alphabet $9.07 billion and Meta $2.12 billion. Microsoft stayed positive, but its free cash flow fell 27% to $18.63 billion, and it said it will go to the bond market to keep building. Meta said on Monday that it plans to raise $95 billion by selling stock.
That is the shift investors priced on Thursday and Friday. A company paying for a bet out of profit can stop when the bet looks bad. A company paying with borrowed money owes the money either way.
The pressure shows most clearly in the smaller companies that rent out GPUs, often called neoclouds. CoreWeave, Nebius and IREN all fell after Microsoft and Alphabet reported. CNBC reports that they need billions in outside capital, lack investment-grade credit ratings, and are watching lenders pull back as interest rates rise. The giants lean on them for capacity, so trouble there turns into delays for everyone. Microsoft’s late GPUs are exactly that kind of delay.
Anthropic’s filing shows the same shape from the other side. Its business lost only $693 million running in the first half, which is close to breaking even for a company growing that fast. But it has agreed to buy $417 billion of computing and data-centre capacity, on contracts that begin in 2026 and 2027 and run up to 20 years. Its revenue has to keep growing for a decade to make those payments comfortable, and nobody can promise that.
China’s selloff is the same story told in another market. Reuters reports that AI shares led the decline after the American selloff revived doubts about the returns on heavy AI spending. The CSI 300 has now lost 15% over 12 straight weeks, the longest losing streak on record. UBS’s head of China internet research, Kenneth Fong, put the mood plainly. He said some stocks are “increasingly treating the spending peak as a permanent earnings burden”.
So, to the question I started with, is the boom fading? On the evidence of this week, I would say the spending is intact and the financing is fragile. That is a different and more ordinary risk than a collapse in demand. Building ahead of revenue with borrowed money works while lenders stay patient and rates stay low. If either of those moves, the companies with the thinnest cushion will feel it first. Those are the neoclouds today, and the labs carrying hundreds of billions in commitments after them.
What would change my mind is a big buyer cutting its plans because the demand was not there, rather than because the chips were late. Nothing like that happened in any of this week’s reports. Watch for it in the first quarter, when the 2027 budgets get set.
